Title: Multi-view Progressive Adaptation for Cross-Domain Few-Shot Segmentation

Abstract:

Cross-Domain Few-Shot Segmentation focuses on identifying object categories within data-scarce target domains, relying on a limited number of reference exemplars. Standard approaches typically first develop few-shot learning capabilities in a large-scale source domain before attempting to adapt these models to target domains. Nevertheless, the scarcity and lack of diversity in target samples often restrict the performance of current methods. Furthermore, the initial deficiency of few-shot proficiency in the target domain, combined with significant domain discrepancies, creates a substantial barrier to effectively leveraging target data and hinders the adaptation process.

To address these challenges, we introduce Multi-view Progressive Adaptation (MPA), a framework that incrementally transfers few-shot capabilities to target domains through both data and strategic enhancements.

(i) Data Perspective: We propose Hybrid Progressive Augmentation, a technique that generates increasingly diverse and complex views by applying cumulative strong augmentations. This process constructs progressively more difficult learning scenarios to enhance model robustness.

(ii) Strategy Perspective: We design Dual-chain Multi-view Prediction, which exploits these complex views via both sequential and parallel learning pathways under comprehensive supervision.

By enforcing prediction consistency across these varied and intricate views, MPA ensures robust and precise adaptation to target domains. Extensive experimental results confirm that MPA successfully transfers few-shot capabilities to new domains, surpassing state-of-the-art methods by a significant margin of +7.0%.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC